Is Distance Matrix Enough for Geometric Deep Learning?

Explore the limitations and potential of geometric deep learning in this comprehensive talk by Zian Li from Valence Labs. Delve into the insufficiency of Message Passing Neural Networks (MPNNs) for learning 3D graph geometry and discover the novel k-DisGNNs approach. Understand how k-DisGNNs can effectively exploit distance matrix information, learn high-order geometric data, unify existing geometric models, and act as universal function approximators. Examine the connection between geometric deep learning and traditional graph representation learning, challenging the notion that complex equivariant models are the only solution. Follow along as Li presents counterexamples, experimental results, and engages in a thought-provoking Q&A session on topics including MD17, rMD17, and QM9 datasets.

- Intro & Overview
- Incompleteness of Vanilla DisGNN
- k-DisGNNs
- Extracting High-Order Geometric Information
- Unifying Invariant Geometric Models
- Completeness and Universality
- Experiments
- Experiments: MD17
- Experiments: rMD17
- Experiments: QM9 and Effectiveness of Edge Repr
- Discussion
- Q+A